Segmentation of drone images

Hi, as you can see in this notebook: Drone images segmentation | Kaggle, mainly obtained by copying and pasting from this tutorial: Fine-Tune a Semantic Segmentation Model with a Custom Dataset
after 100 steps in training, mean Iou begins to drop again, while mean accuracy and overall accuracy are almost constant during training and dont grow . What hyperparameters should I put in for better results? Thank you


I recommend checking out the hyperparameters of the original paper. From that paper:

We trained the models using AdamW optimizer for 160K iterations on ADE20K, Cityscapes, and 80K iterations on COCO-Stuff. (…) We used a batch size of 16 for ADE20K and COCO-Stuff, and a batch size of 8 for Cityscapes. The learning rate was set to an initial value of 0.00006 and then used a “poly” LR schedule with factor 1.0 by default.

Actually with these hyperparameters :

lr = 0.00006
batch_size = 8

I got better results:

(score: 2.932823657989502).
TrainOutput(global_step=80, training_loss=3.012317517399788, metrics={‘train_runtime’: 5060.615, ‘train_samples_per_second’: 0.126, ‘train_steps_per_second’: 0.016, ‘total_flos’: 1.123496604205056e+16, ‘train_loss’: 3.012317517399788, ‘epoch’: 2.0})